A New Framework to Assess the Individual Fairness of Probabilistic Classifiers

Abstract

Fairness in machine learning has become a global concern due to the predominance of ML in automated decision-making systems. In comparison to group fairness, individual fairness, which aspires that similar individuals should be treated similarly, has received limited attention due to some challenges. One major challenge is the availability of a proper metric to evaluate individual fairness, especially for probabilistic classifiers. In this study, we propose a framework PCIndFair to assess the individual fairness of probabilistic classifiers. Unlike current individual fairness measures, our framework considers probability distribution rather than the final classification outcome, which is suitable for capturing the dynamic of probabilistic classifiers, e.g., neural networks. We perform extensive experiments on four standard datasets and discuss the practical benefits of the framework. This study can be helpful for machine learning researchers and practitioners flexibly assess their models' individual fairness. The complete code of the framework is publicly available1.